Why this is in the vault
A concrete, step-by-step Claude Skill that chains two MCP connectors (Notion read, Kit write) into a 7-step email-admin pipeline with a deliberate human checkpoint before the one irreversible action (hitting send) — a real, reusable automation pattern wrapped around a cohort-launch promo, not a pitch disguised as content.
The core argument
Cole (Write With AI co-founder) wrote 70 launch emails in 3 weeks for a Ship 30 for 30 cohort relaunch and found the writing was only half the job — each email also needed spell/grammar check, template selection, loading into Kit (their ESP), segment targeting, a second QC pass, and scheduling. He built a Claude Skill to own everything except the writing and the final send. The pipeline, given a Notion link:
- Spell/grammar check — Claude reads the Notion page, runs the check, and returns a numbered table of errors (type + suggested fix) plus a separate list of softer style notes, rather than auto-correcting. The human picks which numbers to fix; only those get edited.
- Template selection — Claude searches the connected Kit account by template name and confirms the match (carries over styling) rather than requiring the exact template ID.
- Load into Kit as a draft — Claude extracts subject line, preview text, and body copy from anywhere on the Notion page (position-agnostic) and builds the draft with formatting intact; asks before loading if anything is ambiguous.
- Segment/audience targeting — Claude finds the named tag/segment in Kit, confirms it with the human, and sets it on the draft. One manual step remains by design: excluding cold subscribers, which the Skill reminds the human to add rather than doing itself.
- Human QC gate — Claude hands over the Kit draft link and stops. The human reads, edits, uploads images (the Kit MCP can't pull images from Notion), and says "I've QC'd it" to resume.
- Bidirectional sync on final edits — if the human tweaks the draft inside Kit itself, Claude re-runs the error-catch pass and writes any accepted fixes back to both Kit and the Notion source, so the two never drift into different versions.
- Scheduling is the one click never delegated — Cole explicitly keeps "hitting schedule" as human-only.
The piece includes a working, downloadable "Notion to Kit Loader" Skill, built genuinely generic (asks which template/segment each time) with an explicit callout that you can ask Claude to hardcode it to your own template/segment by reading your Kit account — i.e., the Skill self-configures from live account state rather than needing manual editing.
Mapping against Ray Data Co
Strong, direct mapping — this is RDCO's own skills-over-commands practice (feedback_skills_over_commands) applied to a different connector pair, and its QC gate is structurally identical to two hard constraints already in place here: feedback_no_autonomous_external_email (Ray drafts, founder sends — the send-button is human-gated) and feedback_announce_approval_gated_calls. Cole's "hand over the link, stop, wait for 'I've QC'd it'" step is the same shape as Ray's iMessage-approval pattern for anything that crosses an irreversible boundary — the automation owns everything up to the point of no return and explicitly refuses to own that point.
Two things worth taking rather than just nodding at:
- The numbered-table correction UI is better than how
/process-newsletteror vault-write edits typically surface changes. Presenting proposed fixes as a reviewable, selectively-acceptable list (not a silent auto-apply, not a wall of diffs) is a pattern worth reusing anywhere Claude proposes edits to founder-owned prose — it's a lighter-weight version of the fresh-eyes-critic review gate for a different risk tier (copy edits, not strategic output). - Bidirectional sync on step 7 (fix in both Kit and Notion so they never diverge) is the same failure mode RDCO's own Obsidian-sync boundary work guards against (
feedback_obsidian_sync_boundary_deletes) — two copies of the same artifact drifting apart because an edit only landed in one place. Cole's answer (always write back to the source of truth, not just the destination) is the correct instinct and matches the vault's own single-source-of-truth discipline.
Nothing here is new capability RDCO lacks — the Mac Mini agent already does multi-connector chains with human-gated terminal actions — but it's a clean, legible worked example to point external collaborators to when explaining why the approval-gate pattern exists, and the numbered-table correction UI is a concrete upgrade worth stealing for lower-stakes edit proposals.
⚠️ Sponsorship
Self-promo, no third-party sponsor. The email opens by naming Write With AI as "our paid newsletter," pitches "AI Writing Skool" (live sprints/bootcamps/cohorts/Monday Hot Seats) in the second paragraph, and closes with a five-property owned-brand footer (Write With AI, Ship 30 for 30, Premium Ghostwriting Academy, Typeshare, Ghostbase). The Skill itself is given away free as a lead magnet into that ecosystem. None of this disqualifies the content — the 7-step workflow and downloadable Skill are genuinely usable independent of the pitch — but read the framing as acquisition funnel, not neutral tutorial.
Related
- [[2026-04-26-write-with-ai-build-claude-skill-from-scratch]] — the craft thesis ("great prompts are education products") behind why this sender keeps shipping Skills as lead magnets
- [[2026-01-07-write-with-ai-headline-claude-skill]] — an earlier Write With AI Claude Skill from the same stable, same give-away-the-skill pattern
- [[feedback_no_autonomous_external_email]] — the identical human-gated-send constraint already standing in RDCO's own practice
- [[feedback_skills_over_commands]] — the architectural preference this piece is an instance of